blume 2.0.3 → 2.1.0

package/src/ai/index.ts

@@ -20,23 +20,35 @@
  * nothing here runs at request time.
  */
 export {
+  anthropic,
   gateway,
+  gemini,
+  grok,
   inkeep,
   llmgateway,
+  openai,
   openaiCompatible,
   openrouter,
 } from "./ask.ts";
 export type {
   AssistantAdapter,
   AssistantAdapterOptions,
+  AssistantAnthropicAdapter,
+  AssistantAnthropicOptions,
   AssistantGatewayAdapter,
   AssistantGatewayOptions,
+  AssistantGeminiAdapter,
+  AssistantGeminiOptions,
+  AssistantGrokAdapter,
+  AssistantGrokOptions,
   AssistantInkeepAdapter,
   AssistantInkeepOptions,
   AssistantLlmGatewayAdapter,
   AssistantLlmGatewayOptions,
+  AssistantOpenAIAdapter,
   AssistantOpenAICompatibleAdapter,
   AssistantOpenAICompatibleOptions,
+  AssistantOpenAIOptions,
   AssistantOpenRouterAdapter,
   AssistantOpenRouterOptions,
   AssistantProviderOptions,

package/src/ai/llms.ts

@@ -12,7 +12,11 @@ import { API_PAGES_PATH, OPENAPI_PATH } from "./api/paths.ts";
 import { downlevelComponents } from "./component-markdown.ts";
 import { relativeLinkRewriter } from "./relative-links.ts";
 import { projectComponentSerializers } from "./serializers.ts";
-import { AGENT_SKILLS_DIR, AGENT_SKILLS_INDEX_PATH } from "./skills.ts";
+import {
+  AGENT_SKILLS_DIR,
+  AGENT_SKILLS_INDEX_PATH,
+  servesSkillsIndex,
+} from "./skills.ts";
 import type { SkillArtifact } from "./skills.ts";
 import { applyAgentVisibility } from "./visibility.ts";
 
@@ -75,7 +79,7 @@ const agentResourceLines = (project: BlumeProject): string[] => {
       `- [MCP server](${url(config.agents.mcp.route)}): Streamable HTTP Model Context Protocol server with search_docs, get_page, list_pages, and get_navigation tools, plus every page as a resource. Discovery document: ${url("/.well-known/mcp.json")}`
     );
   }
-  if (config.agents.skills) {
+  if (servesSkillsIndex(config)) {
     lines.push(
       `- [Agent skills](${url(AGENT_SKILLS_INDEX_PATH)}): Agent Skills discovery index of the skills this site publishes.`
     );

package/src/ai/markdown.ts

@@ -14,6 +14,7 @@ import { downlevelComponents } from "./component-markdown.ts";
 import { buildLlmsIndex } from "./llms.ts";
 import { relativeLinkRewriter } from "./relative-links.ts";
 import { projectComponentSerializers } from "./serializers.ts";
+import { servesSkillsIndex } from "./skills.ts";
 import { applyAgentVisibility } from "./visibility.ts";
 
 /** One route's raw-Markdown variants. */
@@ -115,7 +116,7 @@ export const buildRawMarkdown = async (
         config: advertisedConfig(project.config, {
           mcp: servesMcp(project, userPages),
           // Skills are collected at the end of the build, after this runs.
-          skills: Boolean(project.config.agents.skills),
+          skills: servesSkillsIndex(project.config),
         }),
       }),
     };

package/src/ai/mcp/data.ts

@@ -130,6 +130,12 @@ export const buildMcpData = async (project: BlumeProject): Promise<McpData> => {
       if (doc.facets) {
         document.facets = doc.facets;
       }
+      if (doc.boost !== undefined) {
+        document.boost = doc.boost;
+      }
+      if (doc.keywords) {
+        document.keywords = doc.keywords;
+      }
       return document;
     }),
     instructions: config.agents.mcp.instructions,

package/src/ai/openapi-components.ts

@@ -134,9 +134,15 @@ export const openapiComponentSerializers = (
         return null;
       }
       const signature = inlineCode(operationSignature(data, operation));
-      return operation.deprecated
-        ? `${signature}\n\n**Deprecated.**`
-        : signature;
+      const notes = [
+        // An agent reading the page must not take a webhook for an endpoint
+        // it can call.
+        operation.webhook
+          ? "**Webhook.** The API sends this request to your endpoint."
+          : "",
+        operation.deprecated ? "**Deprecated.**" : "",
+      ].filter(Boolean);
+      return [signature, ...notes].join("\n\n");
     },
   } satisfies Record<string, ComponentMarkdown>;
 };

package/src/ai/site-skill.ts

@@ -0,0 +1,274 @@
+import { hasGeneratedChangelog } from "../astro/pages.ts";
+import {
+  mountBasePath,
+  normalizeBasePath,
+  withBasePath,
+} from "../core/base-path.ts";
+import { CHANGELOG_INDEX_ROUTE } from "../core/changelog-index.ts";
+import { discoverPagesSync } from "../core/custom-pages.ts";
+import type { BlumeProject } from "../core/project-graph.ts";
+import { absoluteUrl } from "../core/site-url.ts";
+import type { NavNode, PageRecord } from "../core/types.ts";
+import { resolveReferences } from "../openapi/references.ts";
+import { docsPhrase } from "./ai-catalog.ts";
+import { API_PAGES_PATH } from "./api/paths.ts";
+import { asciiSlugify } from "./mcp/discovery.ts";
+import {
+  SKILL_DESCRIPTION_MAX,
+  generatesSiteSkill,
+  skillMdArtifact,
+} from "./skills.ts";
+import type { SkillArtifact } from "./skills.ts";
+
+/**
+ * The site's own agent skill (`agents.skillMd`): a `SKILL.md` in the Agent
+ * Skills format, built from what the build already knows — the site's title
+ * and description, how to read any page as Markdown, llms.txt and the MCP
+ * server, the API references, and a map of the docs with each page's one-line
+ * description. No model writes it, so it costs nothing and is current with
+ * every build: a guide to the docs rather than a summary of the product. A
+ * hand-written skill of the same name in `agents.skills` replaces it.
+ */
+
+/** The most pages the map lists; llms.txt holds the rest. */
+export const MAX_SKILL_PAGES = 200;
+
+/** The Agent Skills spec caps a skill name at 64 characters. */
+const NAME_MAX = 64;
+
+/** Where the site skill is served besides the discovery index. */
+export const SKILL_MD_PATH = "/skill.md";
+
+/**
+ * The site skill's name: the title as a skill name (lowercase alphanumerics
+ * and single hyphens), else the site's host, for a title with no Latin
+ * letters in it.
+ */
+export const siteSkillName = (project: BlumeProject): string => {
+  const { site } = project.config.deployment.options;
+  const fromTitle = asciiSlugify(project.config.title);
+  const name =
+    fromTitle || asciiSlugify(site ? new URL(site).hostname : "") || "docs";
+  return name.slice(0, NAME_MAX).replace(/-+$/u, "");
+};
+
+/** A title as Markdown link text (brackets escaped, as in llms.txt). */
+const linkText = (title: string): string =>
+  title.replaceAll(/[\\[\]]/gu, String.raw`\$&`);
+
+/**
+ * Current-version, default-locale pages worth listing: not API operations,
+ * and not changelog entries, which the changelog index covers in one line.
+ */
+const mappedPages = (project: BlumeProject): PageRecord[] => {
+  const defaultLocale = project.config.i18n?.defaultLocale;
+  return project.graph.pages.filter(
+    (page) =>
+      !(
+        page.meta.ai.exclude ||
+        page.meta.draft ||
+        page.meta.sidebar.hidden ||
+        page.meta.seo.noindex ||
+        page.version !== "" ||
+        page.source.name === "openapi" ||
+        page.contentType === "changelog"
+      ) &&
+      (defaultLocale === undefined || page.locale === defaultLocale)
+  );
+};
+
+/** Renders the page map, listing each page once, up to the cap. */
+class PageMap {
+  readonly #byRoute: Map<string, PageRecord>;
+  readonly #listed = new Set<string>();
+  readonly #url: (route: string) => string;
+
+  constructor(pages: PageRecord[], url: (route: string) => string) {
+    this.#byRoute = new Map(pages.map((page) => [page.route, page]));
+    this.#url = url;
+  }
+
+  /** How many listable pages the cap left out. */
+  get omitted(): number {
+    return this.#byRoute.size - this.#listed.size;
+  }
+
+  /** The Markdown link line for a route, or nothing when it isn't listed. */
+  line(route: string): string[] {
+    const page = this.#byRoute.get(route);
+    if (
+      !page ||
+      this.#listed.has(route) ||
+      this.#listed.size >= MAX_SKILL_PAGES
+    ) {
+      return [];
+    }
+    this.#listed.add(route);
+    const summary = page.description ? `: ${page.description}` : "";
+    return [`- [${linkText(page.title)}](${this.#url(route)})${summary}`];
+  }
+
+  /** One navigation level: its loose pages, then each group under a heading. */
+  level(nodes: NavNode[], depth: number): string[] {
+    const list: string[] = [];
+    const groups: string[] = [];
+    for (const node of nodes) {
+      if (node.kind === "page") {
+        list.push(...this.line(node.route));
+        continue;
+      }
+      const blocks = [
+        ...(node.route ? this.line(node.route) : []),
+        ...this.level(node.children, depth + 1),
+      ];
+      if (blocks.length > 0) {
+        groups.push(
+          "",
+          `${"#".repeat(Math.min(depth, 6))} ${node.label}`,
+          "",
+          ...blocks
+        );
+      }
+    }
+    return [...list, ...groups];
+  }
+
+  /** Pages the navigation doesn't reach, in route order. */
+  rest(): string[] {
+    return [...this.#byRoute.keys()]
+      .toSorted((a, b) => a.localeCompare(b))
+      .flatMap((route) => this.line(route));
+  }
+}
+
+/** The "Reading the docs" bullets: every agent surface the site serves. */
+const surfaceLines = (
+  project: BlumeProject,
+  url: (path: string) => string
+): string[] => {
+  const { agents } = project.config;
+  const lines = [
+    `- Any page as Markdown: add \`.md\` to its URL (the home page is ${url("/index.md")}). Every page link below already points at the Markdown.`,
+  ];
+  if (agents.llmsTxt.enabled) {
+    lines.push(
+      `- [llms.txt](${url("/llms.txt")}): every page with a one-line summary.`,
+      `- [llms-full.txt](${url("/llms-full.txt")}): every page in one file, for when you need all of it.`
+    );
+  }
+  if (agents.mcp.enabled) {
+    lines.push(
+      `- MCP server at ${url(agents.mcp.route)}, with \`search_docs\`, \`get_page\`, \`list_pages\`, and \`get_navigation\`. Connect it to search the docs instead of guessing URLs.`
+    );
+  }
+  if (agents.api) {
+    lines.push(
+      `- [JSON API](${url(API_PAGES_PATH)}): the page index, with each page as JSON and Markdown.`
+    );
+  }
+  const userPages = project.context.pagesRoot
+    ? discoverPagesSync(project.context.pagesRoot)
+    : [];
+  if (hasGeneratedChangelog(project, userPages)) {
+    lines.push(
+      `- [Changelog](${url(`${CHANGELOG_INDEX_ROUTE}.md`)}): release notes, newest first.`
+    );
+  }
+  return lines;
+};
+
+/** The "API reference" bullets, one per rendered reference. */
+const referenceLines = (
+  project: BlumeProject,
+  url: (path: string) => string
+): string[] =>
+  resolveReferences(project.config).map((reference) => {
+    // Scalar renders its own page with no Markdown mirror; Blume-rendered
+    // references are pages like any other, one per operation.
+    const route =
+      reference.kind === "scalar"
+        ? reference.route
+        : withBasePath(reference.basePath, reference.route);
+    const pages =
+      reference.kind === "scalar"
+        ? ""
+        : ", with a page per operation (add `.md` for its Markdown)";
+    return `- [${linkText(reference.label)}](${url(route)})${pages}.`;
+  });
+
+/** The skill's frontmatter `description`: what it's for and when to use it. */
+const skillDescription = (project: BlumeProject): string => {
+  const { description, title } = project.config;
+  const summary = description ? ` ${description.trim()}` : "";
+  return `Read ${docsPhrase(title, "docs")} instead of answering from memory.${summary} Use when a task involves ${title}: how it works, how to set it up or configure it, or its API.`.slice(
+    0,
+    SKILL_DESCRIPTION_MAX
+  );
+};
+
+/** The site skill's `SKILL.md`. */
+const skillMarkdown = (project: BlumeProject, name: string): string => {
+  const { config } = project;
+  const { site = "" } = config.deployment.options;
+  const base = normalizeBasePath(config.deployment.options.base);
+  // Mounted and encoded like llms.txt's links.
+  const url = (path: string): string =>
+    encodeURI(absoluteUrl(site, mountBasePath(base, path)));
+  const markdownUrl = (route: string): string =>
+    url(route === "/" ? "/index.md" : `${route}.md`);
+
+  const map = new PageMap(mappedPages(project), markdownUrl);
+  const pages = [
+    ...map.level(project.graph.navigation.sidebar, 3),
+    ...map.rest(),
+  ];
+  const { omitted } = map;
+  if (omitted > 0) {
+    pages.push(
+      "",
+      config.agents.llmsTxt.enabled
+        ? `…and ${omitted} more. [llms.txt](${url("/llms.txt")}) lists every page.`
+        : `…and ${omitted} more.`
+    );
+  }
+  const references = referenceLines(project, url);
+
+  return [
+    "---",
+    `name: ${name}`,
+    `description: ${JSON.stringify(skillDescription(project))}`,
+    "---",
+    "",
+    `# ${config.title}`,
+    "",
+    ...(config.description ? [config.description.trim(), ""] : []),
+    `This skill points you at ${docsPhrase(config.title, "docs")}, at ${url("/")}. When a task involves ${config.title}, read the relevant page before relying on what you remember: the docs are the source of truth, and they change.`,
+    "",
+    "## Reading the docs",
+    "",
+    ...surfaceLines(project, url),
+    ...(references.length > 0
+      ? ["", "## API reference", "", ...references]
+      : []),
+    ...(pages.length > 0 ? ["", "## Pages", "", ...pages] : []),
+    "",
+  ]
+    .join("\n")
+    .replaceAll(/\n{3,}/gu, "\n\n");
+};
+
+/**
+ * The generated site skill, or `null` when `agents.skillMd` is off or there
+ * is no `deployment.site` to make its links absolute.
+ */
+export const buildSiteSkill = (project: BlumeProject): SkillArtifact | null => {
+  if (!generatesSiteSkill(project.config)) {
+    return null;
+  }
+  const name = siteSkillName(project);
+  return skillMdArtifact(
+    name,
+    skillDescription(project),
+    skillMarkdown(project, name)
+  );
+};

package/src/ai/skills.ts

@@ -29,7 +29,7 @@ const AGENT_SKILLS_SCHEMA =
 const SKILL_NAME = /^[a-z0-9]+(?:-[a-z0-9]+)*$/u;
 const SKILL_NAME_MAX = 64;
 /** The Agent Skills spec caps `description` at 1024 characters. */
-const DESCRIPTION_MAX = 1024;
+export const SKILL_DESCRIPTION_MAX = 1024;
 
 /** One publishable skill artifact plus its index entry. */
 export interface SkillArtifact {
@@ -142,7 +142,7 @@ const skillMeta = (raw: string, dirName: string): SkillMetaResult => {
     };
   }
   return {
-    meta: { description: description.slice(0, DESCRIPTION_MAX), name },
+    meta: { description: description.slice(0, SKILL_DESCRIPTION_MAX), name },
   };
 };
 
@@ -206,6 +206,39 @@ export const collectSkills = async (dir: string): Promise<CollectedSkills> => {
   return { skills, warnings };
 };
 
+/** A skill that is one `SKILL.md`, published verbatim. */
+export const skillMdArtifact = (
+  name: string,
+  description: string,
+  markdown: string
+): SkillArtifact => {
+  const content = new TextEncoder().encode(markdown);
+  return {
+    content,
+    description: description.slice(0, SKILL_DESCRIPTION_MAX),
+    digest: sha256(content),
+    name,
+    path: `${name}/SKILL.md`,
+    type: "skill-md",
+  };
+};
+
+/**
+ * Whether the build generates the site's own skill (`agents.skillMd`): on,
+ * and with a `deployment.site` to make its links absolute, since an installed
+ * skill is read away from the site.
+ */
+export const generatesSiteSkill = (config: ResolvedConfig): boolean =>
+  config.agents.skillMd && Boolean(config.deployment.options.site);
+
+/**
+ * Whether a skills discovery index is served: from `agents.skills`, or for
+ * the generated site skill alone. On an advertised config (see
+ * `advertisedConfig`), both are already switched off when nothing published.
+ */
+export const servesSkillsIndex = (config: ResolvedConfig): boolean =>
+  Boolean(config.agents.skills) || generatesSiteSkill(config);
+
 /**
  * The discovery index (v0.2.0 schema). Artifact URLs are path-absolute under
  * `deployment.base` — the RFC resolves them against the index origin.

package/src/ai/write-skill.ts

@@ -0,0 +1,97 @@
+import { existsSync } from "node:fs";
+
+import { join, relative, resolve } from "pathe";
+
+import type { BlumeProject } from "../core/project-graph.ts";
+import { siteSkillName } from "./site-skill.ts";
+
+/**
+ * `blume skill`'s logic, kept out of the command module so it runs (and is
+ * covered) in-process. Writing a product skill is judgment work — what the
+ * product is, which tasks matter, where it bites — so it lives in the
+ * `blume-write-skill` skill, shipped inside the package. This module only
+ * works out where the skill goes and writes the prompt that points a coding
+ * agent at the skill and the docs.
+ */
+
+/** The skill's directory inside the published package. */
+export const WRITE_SKILL_DIR = join("skills", "blume-write-skill");
+
+/** Where `agents.skills` points when the config doesn't set it. */
+const DEFAULT_SKILLS_DIR = "skills";
+
+/** Where the skill goes, and what the agent needs to know to write it. */
+export interface SkillPlan {
+  /** The `agents.skills` value to add, when the config has none. */
+  configure?: string;
+  /** The config file to add it to, relative to the project root. */
+  configFile: string;
+  /** The content root, relative to the project root. */
+  contentRoot: string;
+  /** Whether a skill is already there to update. */
+  exists: boolean;
+  /** Where to write the skill, relative to the project root. */
+  file: string;
+  /** The skill's name: the generated site skill's, so it replaces it. */
+  name: string;
+  /** The site's URL, for the skill's absolute links. */
+  site?: string;
+  title: string;
+}
+
+/** Plan the skill for a scanned project. */
+export const planSkill = (project: BlumeProject): SkillPlan => {
+  const { config, context } = project;
+  const name = siteSkillName(project);
+  const file = resolve(
+    context.root,
+    config.agents.skills ?? DEFAULT_SKILLS_DIR,
+    name,
+    "SKILL.md"
+  );
+  const plan: SkillPlan = {
+    configFile: context.configFile
+      ? relative(context.root, context.configFile)
+      : "blume.config.ts",
+    contentRoot: relative(context.root, context.contentRoot) || ".",
+    exists: existsSync(file),
+    file: relative(context.root, file),
+    name,
+    site: config.deployment.options.site,
+    title: config.title,
+  };
+  return config.agents.skills
+    ? plan
+    : { ...plan, configure: `./${DEFAULT_SKILLS_DIR}` };
+};
+
+/** The prompt that hands the plan to a coding agent. */
+export const writeSkillPrompt = (
+  plan: SkillPlan,
+  options: { skillDir: string; version: string }
+): string => {
+  const { skillDir, version } = options;
+  const lines = [
+    `Write the agent skill for the ${plan.title} docs, a Blume ${version} site.`,
+    "",
+    `Follow the blume-write-skill skill: read ${join(skillDir, "SKILL.md")} first and work through its workflow.`,
+    "",
+    plan.exists
+      ? `A skill already exists at ${plan.file}. Update it against the current docs rather than starting over, and keep what's still true.`
+      : `Write it to ${plan.file}.`,
+    `Its frontmatter \`name\` must be \`${plan.name}\`: that's the name of the skill Blume generates for this site, and a skill with the same name replaces it.`,
+  ];
+  if (plan.configure) {
+    lines.push(
+      `Then set \`agents.skills: "${plan.configure}"\` in ${plan.configFile} so the build publishes it.`
+    );
+  }
+  lines.push(
+    "",
+    `The docs source is in ${plan.contentRoot}.`,
+    plan.site
+      ? `The site is served at ${plan.site}, so link pages as ${plan.site.replace(/\/$/u, "")}<route>.md.`
+      : `The config has no \`deployment.site\`, and the skill's links must be absolute. Ask me for the site's URL, and set it as \`deployment.site\` in ${plan.configFile}.`
+  );
+  return lines.join("\n");
+};